Starting from where you left off in the search

[paul188]

I am currently trying to use Symbolic Regression as part of an optimization routine.
Now, since I expect the result to be very often very similar from one optimization step to the next, I would like to be able to supply an initial guess or starting point to the Symbolic Regression in order to speed up things.
Is that already possible ? I could not find any such Option in the EquationSearch.

P.S.: And maybe it would also be possible to suppress printing out the results in each iteration.

[MilesCranmer]

You can do warm starts by passing the saved_state option to equation_search. Make sure to specify return_state=true so that equation_search actually returns the entire state the first time.

For example:

saved_state = equation_search(X, y; options=options, return_state=true)

# Continue where it left off:
saved_state = equation_search(X, y; options=options, return_state=true, saved_state=saved_state)

However I might recommend using the MLJ interface which does this automatically. The second time you run fit!(mach), it will propagate all expressions forward. (And will recompute the losses based on whatever the current dataset is).

Also, in case you were also asking about this, the most relevant other threads:

• Start search from a base function #418
• Create Hall of fame manually #443

Other threads that are relevant:

• Specific functional solutions #557
• Make PySR look for strictly factorised functions #465
• Symbolic integrals with SymbolicRegression.jl #401
• How to compute the complexity of a function inside the custom objective function? #380
• [Feature]: Add complexity calculation for user defined expression #339
• How to tell Symbolic Regression to create trees without using constants? #417

---

P.S.: And maybe it would also be possible to suppress printing out the results in each iteration.

You can set this with verbosity=0. See https://astroautomata.com/SymbolicRegression.jl/dev/api/ (though there are a lot of options; apologies :))

[paul188]

Thanks alot ! That is extremely helpful !